Performance Optimization of Photonic Accelerators in Neuromorphic Computing via Structure Trimming

Due to the low latency and parallelism of light, photonic neuromorphic accelerators are expected to realize high-speed computing with low energy consumption, thus meeting the exponentially growing computing demands brought about by the continuous development of artificial intelligence. However, the optical response of many photonic structures used in photonic accelerators depends on their geometry and is thus sensitive to fabrication errors. In this paper, silicon ion implantation, as an emerging post-trimming method, is used to effectively tune the optical response of silicon nitride ring resonators and asymmetric directional couplers. Silicon ion implantation locally alters the material and thus the effective refractive index of the guided mode. Due to the properties of silicon, silicon ion implantation causes negligible loss to the structures at an operating wavelength of 1550 nm, and the trimming is non-volatile. After implantation, ring resonator transmission spectra can be shifted from 6 pm to several nanometres. Further, the transmission change of the asymmetric directional coupler was 17 dB. Silicon ion implantation provides an excellent platform for the non-volatile tuning of photonic structures, facilitating the scaling of photonic neuromorphic accelerators and ultimately advancing artificial intelligence-based applications.